Productivity and efficiency analysis have gained substantial attention in many industries over the last two decades, and stochastic frontier analysis has been one of the most popular analytical approaches. The abundant model choices in stochastic frontier analysis make it difficult to select the best option and compare studies. The main purpose of this study is to examine the sensitivity of technical and scale efficiency estimates to choices around input-output combinations, functional forms, distributional assumptions and estimation methods in stochastic frontier analysis, using data from an Australian fishery to illustrate these effects. We estimated 252 stochastic frontier models using combinations of variable choice, functional form and distributional assumptions. A second stage analysis was conducted to examine the effects of model choices on statistical properties of technical and scale efficiency. The results show that estimates of technical and scale efficiency are most sensitive to distributional assumptions and the choice of time effects. In particular, the assumption of time-invariant efficiency produced significantly higher technical efficiency (20 percentage points) and scale efficiency (8 percentage points) scores than time-varying efficiency models in our analysis. We also find that the choice of fixed input variables can significantly affect the average efficiency estimates, by as much as 5 percentage points, but mean efficiency was not significantly affected by the choice of variable inputs. Our findings suggest that caution should be taken when comparing findings of stochastic frontier studies using different distributional and fixed input assumptions.

在过去的二十年中，生产力和效率分析已在许多行业中引起了广泛关注，随机前沿分析已成为最受欢迎的分析方法之一。随机前沿分析中的大量模型选择使选择最佳方案和比较研究变得困难。这项研究的主要目的是使用澳大利亚渔业的数据来说明技术和规模效率估算对随机前沿分析中投入产出组合，功能形式，分布假设和估算方法周围选择的敏感性。我们使用变量选择，功能形式和分布假设的组合估计了252个随机前沿模型。进行了第二阶段分析，以检查模型选择对技术和规模效率的统计特性的影响。结果表明，技术效率和规模效率的估计对分布假设和时间影响的选择最敏感。特别是，在我们的分析中，时不变效率的假设产生的技术效率（20个百分点）和规模效率（8个百分点）得分明显高于时变效率模型。我们还发现，固定输入变量的选择会显着影响平均效率估计值，最高可达5个百分点，但平均效率并未受到变量输入的选择的显着影响。